ddpm_cifar

→DDPMModel

ddpm_cifar(pretrained: bool | str = False, weights: DDPMCifarWeights | None = None, overrides: object = {})

source edit

Construct a DDPM U-Net trunk for the CIFAR-10 setup.

Paper-faithful CIFAR-10 configuration from Ho, Jain, and Abbeel, 2020 Appendix B.1 / Table 9: sample size 32x32, base_channels=128, channel_mult=(1, 2, 2, 2) (four spatial stages — 32 -> 16 -> 8 -> 4), 2 ResBlocks per stage, self-attention at the 16x16 feature map, 1 attention head, dropout 0.1, and a 1000-step linear beta schedule.

Model Size

Parameters

pretrainedbool or str= False

Weight selector. False → random init; True → the official google/ddpm-cifar10-32 checkpoint (DDPMCifarWeights.CIFAR10) — the trained noise predictor

\epsilon_\theta(x_t, t)

weights(DDPMCifarWeights, optional, keyword - only)= None

Explicit weights enum member; takes precedence over pretrained.

**overridesobject= {}

Optional DDPMConfig field overrides (e.g. dropout=..., num_train_timesteps=...) forwarded into the underlying config. Overrides that change the U-Net shape are incompatible with the pretrained checkpoint.

Returns

DDPMModel

Bare U-Net trunk configured with the CIFAR-10 setup (pretrained when requested).

Notes

Reference: Ho, Jain, and Abbeel, "Denoising Diffusion Probabilistic Models", NeurIPS, 2020 (arXiv:2006.11239), Appendix B.1.

Examples

>>> import lucid
>>> from lucid.models.generative.ddpm import ddpm_cifar
>>> model = ddpm_cifar().eval()
>>> x_t = lucid.randn((1, 3, 32, 32))
>>> t = lucid.tensor([42]).long()
>>> out = model(x_t, t)
>>> out.sample.shape   # (1, 3, 32, 32) — predicted noise
(1, 3, 32, 32)

Used by 1

lucid.models.generative.ddpm

>>> import lucid >>> from lucid.models.generative.ddpm import ddpm_cifar >>> model = ddpm_cifar().eval() >>> x_t = lucid.randn((1, 3, 32, 32)) >>> t = lucid.tensor([42]).long() >>> out = model(x_t, t) >>> out.sample.shape # (1, 3, 32, 32) — predicted noise (1, 3, 32, 32)